Instructions for running the models

In order to use the model code WinBUGS1.4 must be installed for unrestricted 
use. The program and the activation key can be downloaded free of charge from 
http://www.mrc-bsu.cam.ac.uk/bugs/winbugs/contents.shtml.


The models can be run directly using WinBUGS by entering data in WinBUGS format. 
Because this can be time consuming we have made available an interface to 
run the models from the popular open source system for statistical computation 
and graphics, R. The interface to the models requires the packages coda, 
R2WinBUGS, and tcltk. It was written for R version 2.0.1 but we expect it to run 
using later versions unless major changes are made to the packages on which it 
depends. R has a home page at http://www.R-project.org/. The necessary packages 
can be installed from the Comprehensive R Archive Network (CRAN). As an example 
we include the data set “testbugs.txt”. This data was analyzed in our 
paper. It consists of four columns of counts of numbers of individuals for fifty 
species of butterfly representing four contrasting sites. Non-occurrence is 
represented by zeros. If expert knowledge suggests that a species could not 
possibly be expected to occur at a site then zero abundance can be substituted 
by NA (“Not Available”). The data set can also be padded with zeros for species known to 
be present but not caught at any of the sites during the study period. This 
approach is clearly not suitable for completely unknown communities. It was 
designed for comparatively well known groups of organisms such as butterflies, 
trees and flowering plants for which reasonably comprehensive species lists are 
available. 

Spatial and temporal replication within sites is implicitly, but not explicitly, 
modelled under this approach. The implicit justification is that replication has 
reduced the effects of over-dispersion in the data (clumping of individuals) 
making the assumption of Poisson sampling acceptable. It is the responsibility 
of the researcher to ensure that this is an acceptable assumption. MCMC fitting 
has the advantage that models can be extended to explicitly model over 
dispersion. However this is only practical for larger data sets. We advise that 
the use of these comparatively simple models be restricted to circumstances 
where an estimation of the uncertainty associated with comparisons between sites 
is difficult by other means. 

In order to run the analysis place the model files in a folder together with the 
data. Then source in the R code or copy and paste it into the R console. A 
graphical interface will appear. Check that all the paths are correct. Running 
the models from the interface should then be self explanatory. HPD confidence 
intervals for the parameters are output to the R console. Note that the 
interface will provide sensibly scaled figures for data sets with any number of 
species. However the default boxplots become rather overcrowded if more than 
around six sites are analyzed. Knowledge of the R language is needed in order to 
adapt the analysis to more complex situations. Output from WinBUGS is read into 
the R object “CodOut” and HPD confidence intervals are in the object “cis”.